ing at the occlusion phenomenon in the visual object, we propose a novel occlusion boundary detection approach for deep images based on the spectral clustering. Firstly, a new occlusion-related feature, effective standard deviation feature, is defined. Secondly, some pixels are extracted by using mean chi-square set distance, and the similarity matrix is constructed based on the occlusion-related feature. Thirdly, the Laplacian matrix of all the pixels and approximation eigenvectors are approximated by Nystrom approximation method based on the similarity matrix. Then, the obtained approximation eigenvectors are clustered to divide all the pixels in the depth image into two categories, namely the occlusion boundary points and non-occlusion boundary points. Finally, the occlusion boundary of the depth image is obtained by visualizing occlusion boundary points. Experimental results show that the proposed method which does not need any labeled samples has good effectiveness and generality for occlusion boundary detection of the object in the depth image.